summaryrefslogtreecommitdiffstats
path: root/train_ti.py
Commit message (Expand)AuthorAgeFilesLines
* Added TI decay start offsetVolpeon2023-01-131-2/+8
* Code deduplicationVolpeon2023-01-131-60/+26
* UpdateVolpeon2023-01-121-16/+14
* Fixed TI decayVolpeon2023-01-121-8/+3
* Disable Adam weight decayVolpeon2023-01-121-1/+1
* FixVolpeon2023-01-111-3/+3
* HeckVolpeon2023-01-111-1/+1
* TI: Use grad clipping from LoRA #104Volpeon2023-01-111-8/+11
* Better defaultsVolpeon2023-01-101-4/+4
* FixVolpeon2023-01-101-1/+1
* Added arg to disable tag shufflingVolpeon2023-01-101-6/+16
* Enable buckets for validation, fixed vaildation repeat argVolpeon2023-01-091-4/+1
* Add --valid_set_repeatVolpeon2023-01-091-0/+22
* FixVolpeon2023-01-081-1/+1
* Improved aspect ratio bucketingVolpeon2023-01-081-1/+15
* Fixed aspect ratio bucketing; allow passing token IDs to pipelineVolpeon2023-01-081-11/+13
* Improved aspect ratio bucketingVolpeon2023-01-081-48/+37
* CleanupVolpeon2023-01-071-44/+16
* Made aspect ratio bucketing configurableVolpeon2023-01-071-0/+13
* Added progressive aspect ratio bucketingVolpeon2023-01-071-48/+46
* UpdateVolpeon2023-01-071-3/+3
* Relaxed ema_power for TIVolpeon2023-01-061-1/+1
* UpdateVolpeon2023-01-061-24/+17
* Use context manager for EMA, on_train/eval hooksVolpeon2023-01-061-54/+66
* Add contextmanager to EMAModel to apply weights temporarilyVolpeon2023-01-061-33/+24
* Log EMA decayVolpeon2023-01-051-1/+2
* Added EMA to TIVolpeon2023-01-051-1/+58
* UpdateVolpeon2023-01-051-7/+7
* FixVolpeon2023-01-051-8/+11
* Various cleanupsVolpeon2023-01-051-65/+21
* UpdateVolpeon2023-01-041-1/+1
* Various updatesVolpeon2023-01-041-0/+8
* Better eval generatorVolpeon2023-01-041-6/+6
* Fixed reproducibility, more consistant validationVolpeon2023-01-041-5/+19
* Don't use vector_dropout by defaultVolpeon2023-01-031-1/+1
* Added vector dropoutVolpeon2023-01-031-5/+19
* Fixed LR finderVolpeon2023-01-021-3/+3
* UpdateVolpeon2023-01-021-3/+12
* FixVolpeon2023-01-021-4/+2
* Save args before training, tooVolpeon2023-01-021-0/+4
* FixVolpeon2023-01-021-2/+1
* UpdateVolpeon2023-01-021-6/+6
* Improved one cycle schedulerVolpeon2023-01-021-1/+1
* UpdateVolpeon2023-01-011-1/+1
* UpdatesVolpeon2023-01-011-26/+25
* Fixed accuracy calc, other improvementsVolpeon2023-01-011-3/+33
* UpdateVolpeon2022-12-311-0/+1
* Simplified multi-vector embedding codeVolpeon2022-12-311-3/+2
* FixesVolpeon2022-12-311-2/+1
* Added multi-vector embeddingsVolpeon2022-12-311-44/+44